<p>Deep learning stereo matching faces catastrophic accuracy decline in cross-domain scenarios due to inter-domain distribution bias. To address this, we propose DFCO-Net, a novel network that integrates dynamic feature fusion and consistency optimization for superior cross-domain generalization. First, we design a local–global feature collaborative encoder (LGF-CoEncoder) that synergizes fine-grained local features with global contextual information via an adaptive modulation process guided by a holistic scene signature, amplifying salient structural cues while suppressing noise. To resolve inherent matching ambiguities, our progressive hierarchical cost volume aggregation (PHCV-Agg) module employs a two-stage fusion strategy and a structural anchor mechanism within its coarse-to-fine framework to enforce geometric consistency, thereby effectively suppressing error accumulation. Finally, we introduce a feature consistency-guided uncertainty-aware disparity optimization mechanism that directly leverages feature matching consistency as an uncertainty metric to precisely identify and apply targeted corrections to noisy matches. Extensive ablation and comparative studies across diverse datasets validate the effectiveness of each proposed module and the overall robustness of our DFCO-Net. For instance, on the challenging synthetic-to-real cross-domain task, our model reduces the error rate on the KITTI 2015 dataset by over 12% compared to the high-performing PCW-Net. Furthermore, the model demonstrates excellent in-domain performance by achieving a competitive accuracy of 0.72 EPE on the Scene Flow benchmark, coupled with an inference speed (0.16s) that is significantly faster than other methods with comparable accuracy.</p>